Creating a Reliable Music Discovery and Recommendation System

Abstract

AbstractThe aim of this chapter is to show problems related to creating a reliable music discovery system. The SYNAT database that contains audio files is used for the purpose of experiments. The files are divided into 22 classes corresponding to music genres with different cardinality. Of utmost importance for a reliable music recommendation system are the assignment of audio files to their appropriate genres and optimum parameterization for music-genre recognition. Hence, the starting point is audio file filtering, which can only be done automatically, but to a limited extent, when based on low-level signal processing features. Therefore, a variety of parameterization techniques are shortly reviewed in the context of their suitability to music retrieval from a large music database. In addition, some significant problems related to choosing an excerpt of audio file for an acoustic analysis and parameterization are pointed out. Then, experiments showing results of searching for songs that bear the greatest resemblance to the song in a given query are presented. In this way music recommendation system may be created that enables to retrieve songs that are similar to each other in terms of their low-level feature description and genre inclusion. The experiments performed also provide basis for more general observations and conclusions.KeywordsMusic information retrievalMusic databasesMusic parameterizationFeature vectorsPrincipal component analysisMusic classification